Field-deployable lightweight YOLOv8n for real-time detection and counting of Maize seedlings using UAV RGB imagery.

The aim of this study is to propose a lightweight YOLOv8n maize seedling detection algorithm that incorporates multi-scale features to address the problems of large number of model parameters and computation, low real-time performance, and small detection range of the existing maize seedling detection models during plant detection. By fusing RepConv with HGNetV2 using the idea of reparameterisation, a Rep_HGBlock structure is designed to form a new lightweight backbone network, Rep_HGNetV2,; BiFPN is introduced into the neck network portion of the model to enhance the interactive fusion of bidirectional information flow between multiple scales and hierarchies; and a fusion task decomposition, dynamic convolutional alignment is designed, DFL (Distribution Focal Loss) ideas, TDADH, a task dynamically aligned detection head, which uses shared convolution and dynamically aligns the tasks of classification and localization to extract features; and Grad-CAM++ technique is used to generate a heat map for model detection, visualize effective features of the target and understand the model focus region. The experimental results show that the improved model achieves a detection accuracy of 96.5%, which is basically the same as the original model. The weight size, number of parameters, and computational FLOPs are reduced to 3.5 MB, 1.58 M, and 7.4 G, respectively, which are reduced by about 43%, 47%, and 8.6%. The frame rate FPS is only reduced from 149.98 to 146.3, a reduction of about 2.4%. The results show that the lightweight model has high recognition accuracy, speed and low complexity, which is more suitable for practical deployment in resource-constrained edge devices, UAVs, and embedded systems, and is able to provide technical support for the precise management of maize during the seedling stage of drip irrigation water-fertilizer integration.